Google unveils 8th-gen TPU 8t and TPU 8i on Cloud Next Day 2: 3x Ironwood, purpose-built for training and agent serving

Google unveiled the eighth generation of Tensor Processing Units at Cloud Next 2026 on April 23, 2026, one day after the Day 1 Ironwood (7th-gen) GA announcement. The 8th-gen ships as two purpose-built variants: TPU 8t for training, TPU 8i for inference.

TPU 8t: the training chip

• 9,600 TPUs per superpod (up from Ironwood’s 9,216).
• 2 PB shared high-bandwidth memory per superpod.
• 3x the processing power of Ironwood at equivalent chip count.
• 2x performance per watt vs 7th-gen.
• Built for frontier-model pretraining and long post-training runs.

TPU 8i: the inference chip

• 1,152 TPUs per pod, tighter fabric for latency.
• 3x on-chip SRAM vs Ironwood for attention-layer acceleration.
• Designed to serve millions of concurrent agents cost-effectively.
• Targets the 2026 agent-serving bottleneck: many small fast queries with shared-context state.

Why two chips

Training and inference stress silicon differently. Ironwood had already split inference out as the design target for 7th-gen. The 8th-gen doubles down: TPU 8t is a scale-up training monster (bigger superpod, wider HBM), and TPU 8i is a scale-out inference chip (more tightly-clustered pod, SRAM-heavy). Hyperscaler customers can now pick the silicon that matches their workload rather than running inference on training-sized boxes.

Ironwood vs. 8t/8i positioning

Ironwood (TPU v7) remains GA today. The 8th-gen is a roadmap reveal with staged availability through 2026-2027. Expect:

• TPU 8i preview: late 2026 via select Google Cloud regions.
• TPU 8t GA: 2027.
• Anthropic’s 1M-TPU commitment (announced Day 1) scales across Ironwood and into 8th-gen as capacity comes online.

Competitive read

• vs Nvidia Blackwell / Vera Rubin: Google takes the perf/watt claim seriously with TPU 8t’s 2x improvement. Raw peak perf vs Nvidia’s latest remains Nvidia’s lead; Google’s bet is total-cost-of-ownership on sustained workloads.
• vs AWS Trainium2/3: AWS’s custom-silicon story is vertically integrated with its Anthropic anchor customer. Google has done the same through the Anthropic TPU commitment.
• vs Azure’s Maia and Cobalt: Microsoft’s custom-silicon program is further behind; Azure still runs most inference on Nvidia.

For Gemini users, the TPU 8t/8i roadmap points to continued downward pressure on Gemini API pricing through 2027 as capacity compounds. Gemini Flash-tier pricing is already near commodity; Pro tiers have runway to fall further.

Related

• Google unveils Ironwood TPU at Cloud Next 2026
• Google assembles four-partner silicon coalition
• Amazon commits up to $25B more to Anthropic